How tingofurro/keep_it_simple works
Keep It Simple is a neural text-simplification system: users provide a paragraph and receive candidate rewrites with insertions and deletions highlighted. Its training workflow samples rewrites, scores them for coverage, simplicity, fluency, and guardrail violations, then updates the generator with reinforcement learning. The repository also includes a human-evaluation study interface and its study data. Connections reflect the README and shown source-index references; unsampled implementation details and interface wiring are not assumed.
Main entry points: train_keep_it_simple.py Training loop · model_generator.py Text generator · run_keep_it_simple.py Run interface · model_salience.py Coverage scorer
Components 18
Simplification
- Run interface
run_keep_it_simple.py - Text generator
model_generator.py - Edit rendering
utils_edits.py
Reward Scoring
- Coverage scorer
model_salience.py - Text masking
utils_masking.py - Fluency scorer
model_fluency.py - Simplicity scorer
model_simplicity.py - Output guardrails
model_guardrails.py - Score aggregation
utils_scoring.py
Model Training
- Training loop
train_keep_it_simple.py - Optimizer utilities
utils_optim.py - Reinforcement utilities
utils_rl.py - Runtime utilities
utils_misc.py - Training timing
utils_timing.py
Human Evaluation
- Study interface
study_interface/simpli_study.html - Study materials
study_interface/simplification_user_study.json
Other components
- Text user
- Study participant
Connections 21
- Text user → Run interface: submits text
README.md - Run interface → Text generator: generates candidates
run_keep_it_simple.py - Run interface → Edit rendering: renders changes
run_keep_it_simple.py - Training loop → Text generator: samples rewrites
train_keep_it_simple.py - Training loop → Score aggregation: scores candidates
train_keep_it_simple.py - Training loop → Coverage scorer: configures reward
train_keep_it_simple.py - Training loop → Fluency scorer: configures reward
train_keep_it_simple.py - Training loop → Simplicity scorer: configures reward
train_keep_it_simple.py - Training loop → Output guardrails: configures penalties
train_keep_it_simple.py - Training loop → Reinforcement utilities: uses updates
train_keep_it_simple.py - Training loop → Optimizer utilities: uses optimization
train_keep_it_simple.py - Training loop → Training timing: records timing
train_keep_it_simple.py - Training loop → Runtime utilities: uses utilities
train_keep_it_simple.py - Text generator → Output guardrails: applies constraints
model_generator.py - Text generator → Text masking: uses masking
model_generator.py - Text generator → Score aggregation: uses scoring
model_generator.py - Text generator → Runtime utilities: uses batching
model_generator.py - Coverage scorer → Text masking: masks source
model_salience.py - Fluency scorer → Optimizer utilities: builds optimizer
model_fluency.py - Study participant → Study interface: takes evaluation
README.md - Study interface → Study materials: uses materials
README.md